Script loading

Complementary test nets pin the script-loading subsystem (design, implementation). The fixture-site nets run in npm run test:repo; the build-time vendoring and browser nets run in npm run test:visual.

Golden net

scripts-golden.test.mjs pins, across fixture configurations:

• the rendered script region of selected pages, and
• every locally built script, byte-exact.

Fixture builds, region extraction, and the comparison form live in lib/scripts-goldens.mjs. The net was committed and verified green against untouched main before the scripts.html decomposition, so the refactor’s no-output-change claim is machine-checked, not asserted.

The goldens are committed. After a reviewed, intended output change, refresh them with:

npm run update:scripts-goldens

Dispatch net

scripts-dispatch.test.mjs pins the dispatcher’s page-flag wiring: the .Page.Store-gated partials (Mermaid, KaTeX) are dispatched on flagged pages only. The real partials fetch remote assets at build time, so fixture marker overrides stand in for them; what’s pinned is exactly the gate-to-partial wiring, offline.

Loop-contract tests

plugins.test.mjs pins the plugin loop’s registry contract:

• Emission: enable/defer handling, deterministic order, env-override booleans, companions and shims (a shim-gated plugin, and a head-end flag widening its gate), SRI in development builds.
• Validation: shape-guard warnings (the pre-release list shape and a site’s own params.docsy included), name and field allowlisting (the _docsy-shim suffix refused, unknown fields warned), the version guard’s warning and error policy.
• Layering: theme plugins through Hugo’s config merge (inheritance, override, turn-off).

Three companion nets pin the conversions:

• tabpane-persist-plugin.test.mjs: the ungated default, persistence opt-out, and theme-plugin shadowing.
• markmap-plugin.test.mjs and click-to-copy-plugin.test.mjs: the per-conversion contracts. The markmap cases stub the vendoring companion with a marker to stay offline; the real vendor fetch is covered by the build-time vendoring net.

Acceptance test

plugins-acceptance.test.mjs proves adoption end to end: a project site drops assets/js/plugins/hello.js plus one registry entry and gets its script loaded, with zero layout overrides asserted structurally (the fixture contains no layouts/ directory).

Build-time vendoring

The MarkMap vendoring net in the tests/visual/ directory uses real build-time CDN fetches, without a browser. It compares bilingual MarkMap builds with single-version controls to verify each language’s published autoloader bytes; distinct URLs alone cannot prove correct resource-cache behavior.

Runtime nets

Two browser nets under tests/visual/:

• js-runtime.test.mjs loads representative fixture pages in a real browser and asserts that no uncaught exception or in-scope console error fires, alongside behavior probes (search, diagrams, navbar, copy button, tab persistence). Markup and visual goldens can’t see JS runtime breakage (a missing global, a botched conversion); this net can (#1436).
  • Two fixture variants cover both search bundles: the main bundle (scripts/main-bundle.html) concatenates offline-search.js or search.js into main.js, never both.
  • Pages load their real CDN script dependencies, so the net needs network access. The console tally filters off-origin and non-code resource noise but keeps same-origin script and stylesheet load failures (a broken first-party bundle is a defect). Filtered breakage that throws (a dependent script’s missing global) still surfaces as a page error; silent feature degradation is what the behavior probes catch.
• plugins-runtime.test.mjs proves an emitted plugin actually executes: its DOM effects land. A static-markup check can bless output whose runtime is broken (a botched build); this net can’t.

Red-proof rationale

A net that passes for the wrong reason (building nothing, matching an empty region) is worse than a red one: it hides breakage behind green. The nets were built red-first, and where a no-op could masquerade as success, a persistent safeguard proves the signal:

• Zero-output cases are asserted against: a golden’s script region must be non-empty.
• The plugin runtime net’s red-proof doubles as an assertion: a deliberately broken plugin must be the error tally’s only entry, so an empty tally from the healthy plugin is meaningful.
• The site runtime net carries a collector self-test: a page that deliberately throws and drops a same-origin script must have both reported, so a silent collector (wrong event names, races, a broken filter) can’t masquerade as all-green.